![[FILE] Austrian Chancellor Karl Nehammer stands in front of a display reading 'AI' as he attends a press statement during a location summit on 'Impulses for Europe as a business location' organized by the Austrian Government at the headquarters of Austrian photovoltaic manufacturer Fronius International GmbH in Pettenbach, Austria, 14 March 2024. Austrian Chancellor Karl Nehammer points up that it is necessary to reduce energy and production costs, strengthen research and innovation and to cut bureaucracy at national and EU level to support European industry and companies in a market competition with the US, China and Latin America. EFE/EPA/CHRISTIAN BRUNA](https://i0.wp.com/efe.com/wp-content/uploads/2026/07/rss-efe9cea9029566dad85e8bca00fe5bdcf54747778b8w.webp?fit=1920%2C1280&ssl=1)
Beijing, July 18 (EFE).- Chinese startup Moonshot AI has unveiled Kimi K3, which it describes as the world’s largest open-source artificial intelligence model, claiming it outperforms several leading systems from OpenAI and Anthropic in key benchmarks amid an intensifying tech rivalry between Beijing and Washington.
Moonshot AI said Kimi K3 features 2.8 trillion parameters, the measure commonly used to indicate an AI model’s complexity during training, surpassing other open-source Chinese models such as DeepSeek V4 Pro (1.6 trillion parameters) and Zhipu AI’s GLM-5 series (744 billion).
According to the company, Kimi K3 is designed for long-context programming, complex reasoning, and knowledge-intensive tasks. The model is already available through the Kimi chatbot, the Kimi Work application, Kimi Code, and its API, while the full release of its model weights is scheduled for July 27.
Competitive performance at a lower cost
Moonshot acknowledged that K3 «still trails the most powerful proprietary models,» namely Claude Fable 5 and GPT-5.6 Sol, but said it consistently outperformed the other models included in its comparisons.
In an evaluation by AI benchmarking platform Arena.AI, K3 ranked first in six of seven categories, including branding and marketing, reference-based design, data and analytics, consumer products, simulations, and content creation tools, and finished second only in video game-related tasks, behind Claude Fable 5.
Consultancy Artificial Analysis rated K3’s overall performance as comparable to Claude Opus 4.8 and GPT-5.5, regarded as Anthropic’s and OpenAI’s second-tier flagship models, respectively, although it still ranked behind Claude Fable 5 and GPT-5.6 Sol.
Artificial Analysis also noted that K3’s higher performance comes at a higher cost. It charges $15 per million output tokens, translating into an average cost of about $0.95 per task.
Although more expensive than other open-weight Chinese models, K3 remains cheaper than GPT-5.6 Sol, at $1.04 per task, and costs nearly half as much as Claude Opus 4.8, at approximately $1.80 per task.
China-US AI rivalry
The launch comes amid intensifying competition among Chinese AI developers, including Moonshot AI, Zhipu, DeepSeek, and Alibaba, which are relying on open-source models, competitive pricing, and rapid commercialization to narrow the gap with US rivals.
It also follows the release of Zhipu’s GLM-5.2 model last month.
Chinese developers continue to face US restrictions on exports of advanced semiconductors and chipmaking equipment. However, Bank of America analysts, cited by CNBC, said Kimi K3 demonstrates that Chinese firms continue to advance despite those curbs.
The announcement coincided with the opening of the World Artificial Intelligence Conference in Shanghai, where Chinese President Xi Jinping described AI as a strategic priority and called for greater support for open-source development so that «all sectors and companies can benefit from AI.» EFE
imh-sk
